Research programmes
Causal systems
Causal policy learning inside institutions.
A research proposal on identifying policy effects in multi-agent institutions through controlled comparisons and intervention histories.
Research proposal
This programme proposes recording the evidence, policy, authority and outcome around institutional decisions. Those histories would support a causal study, but records alone cannot establish which intervention caused an outcome.
This brief sets out a question and proposed method. It does not report experimental findings. Etylon project · Updated 20 October 2026.
Structured intervention histories should let an institution recover policy effects more reliably than retrospective narratives assembled from correlated records.
Research setting
Decisions, interventions and outcomes in one trace.
The proposed trace connects an intervention to its context and outcome, so the assumptions behind a causal interpretation can be examined.
Proposed research protocol
Instrument decisions before attributing effects.
- 01Instrument
Record the evidence, authority and policy state behind each decision.
- 02Intervene
Introduce controlled policy variants inside bounded domains.
- 03Attribute
Estimate which institutional action changed the observed outcome.
- 04Counterfact
Test whether the explanation predicts plausible alternatives.
- 05Transport
Evaluate whether recovered effects survive a change of context.
A case to examine
Suppose an institution changes the priority rule for shared compute. Completion times improve, but the incoming tasks also become easier. The study must distinguish the effect of the rule from the change in workload, rather than attributing every improvement to the intervention.
What to compare
Define the policy effect of interest and compare policy variants under a declared assignment procedure. Randomization may be appropriate within a bounded setting. Because agents share resources, one team’s treatment can affect another team’s outcome; the design must account for that interference. Compare the resulting estimates with explanations based only on observational correlations. Examine whether feedback improves decisions under controlled interventions. Distinguish achieved outcomes from rewarded proxies and account for effects on other participants.
Conditions and limitations
Document assignment, confounding, missing outcomes and uncertainty. Test predictions against held-out interventions where the setting makes them observable. A counterfactual for one individual event is generally not directly observed. Transferring an estimated effect to another institution requires examining which relevant conditions changed.
Hernán and Robins, Causal Inference: What If. Background on causal estimands, identification assumptions and study design. This reference supplies methodology, not evidence of an Etylon result.
Proposed measures
Define the evidence before judging the result.
- Effect recovery
How closely an estimate recovers a known intervention effect in a controlled test environment.
- Counterfactual calibration
Agreement between predicted alternatives and outcomes from comparable held-out interventions.
- Policy attribution
Errors in assigning an observed change to the policy rather than another cause.
- Transport across contexts
Whether effect estimates remain useful after specified changes in tasks or institutional conditions.
- Evidence coverage
The proportion of evaluated decisions with the records required by the study design.
- Explanation stability
Sensitivity of a causal explanation to plausible changes in assumptions or available evidence.
What would count against the proposition
The resulting causal attributions perform no better than correlational baselines, or collapse when the same policy is tested in a neighbouring institutional context.